Abstract:
|
In longitudinal studies, researchers might be interested in a value that summarizes some time-varying measurement: For example, the duration of brain dysfunction during critical illness may be an important predictor of long-term cognitive problems. Particularly as the use of electronic medical record data increases, information required to summarize these outcomes (eg, daily mental status) may be available at some time points and not at others, posing the challenge of how to use all data that is available while reducing the chance that missing data will bias results. In the context of multivariable regression, we compare several approaches to imputing and summarizing partially missing longitudinal data: 1) we assign subjects with incomplete data a missing value for the entire summary outcome; 2) we assume all missing time points have the a) most and b) least extreme value; and 3) we multiply impute missing data at each individual time point, calculate a summary statistic for each imputation, and use these partially imputed summary statistics in multivariable regression with multiple imputation. For each, we discuss statistical performance, interpretation issues, and ease of use.
|